Differentiation grade is an important biologic feature of malignant tumours, frequently incorporated into prognostication and influencing strategic decisions in patient management. Though being routinely identified during histopathological examination of biopsy samples or surgical material, noninvasive identification of tumour grade by means of semiquantitative and quantitative analysis of medical images represents interest in the context of combining radiomics with important clinical data. In the current study, we attempted to incorporate multiple semiquantitative and quantitative metabolic features into multiparametric modelling in order to try to differentiate tumour grades noninvasively.
In our dataset, we used multiparametric analysis by group method of data handling, incorporating all the extracted semiquantitative and quantitative features to create models that appeared to discriminate all three tumour grades in epithelial tumours with an overall accuracy ranging from 71 to 100%. It should be noted that these results were achieved for all four different segmentation technique datasets. The relatively most accurate model was achieved in subset with active contouring technique (ITK-SNAP segmentation).
Tumour segmentation in metabolic PET images is one of the most important steps in radiomics as it defines a group of voxels that are being assigned as representing active tumour tissue and all further mathematical manipulations for extraction of quantitative indices relies on these selected voxels. The segmentation is affected by various factors, both intrinsic and extrinsic, such as spatial resolution, noise level, shape, and location of pathologic tracer uptake [21]. Low spatial resolution of metabolic PET, especially, compared to anatomic imaging modalities, makes it difficult to define the precise tumour borders. Through recent years of research, a wide variety of segmentation or delineation techniques were proposed, including manual, thresholding-based, and boundary-based methods. Unfortunately, still no general agreement exists on optimal segmentation technique for PET radiomic studies.
First of all, different segmentation methods produce different values of quantitative parameters, mostly due to the inclusion or non-inclusion of necrotic tumour portions. The thresholding techniques with different cutoff values (SUV 2.5, relative thresholds of certain percentage of SUVmax, adaptive thresholding, for example Nestle’s method [22]) are ones that are more commonly used due to simplicity and intuitive and rapid workflow. Nevertheless, they are known to underestimate tumour volume and are susceptible to contrast variations, noise levels, and heterogeneity [23, 24].
In the current study, we used segmentation techniques offering different approaches. SUV 2.5 threshold technique was chosen as representing an “everyday practical” approach, as one of the simplest and less time consuming method, being easily incorporated into everyday practice. However, since SUV 2.5 threshold was first introduced in 2001, its clinical value was validated just for the solitary pulmonary nodule scenario [25]. Liver pool threshold technique was chosen in order to try to extend principles implemented in PET/CT imaging in lymphomas, being one of the reference sites in Deauville scoring system [26]. Forty-one percent SUVmax threshold technique was chosen as one of the methods suggested by EANM guidelines [13]. ITK-SNAP algorithm segmentation was chosen as representing an alternative, non-threshold approach, relying on three-dimensional active contour methods.
As theoretically expected, initial analysis of both semiquantitative and quantitative texture features showed that different techniques or thresholds to segment MTV results in significantly different values of this parameter.
It should be noted that there was no significant difference for all three conventional volumetric parameters between SUVmax with 2.5 fixed threshold technique and liver pool fixed threshold technique. This can be explained by the small difference between liver pool values (usually fluctuating between 2 and 3 SUV with our scanner, imaging protocol, and reconstruction algorithm) and SUVmax 2.5 threshold value.
Nevertheless, incorporating all extracted features into multiparametric modelling provided comparable ability of extracted data from different segmentation subsets to predict tumour differentiation grade.
Previously published studies have demonstrated how single semiquantitative PET parameters may be utilised to differentiate tumour grades, for instance in meningiomas [27], by means of tumour to grey matter ratio of 18F-FDG uptake (TGR). The TGR in high-grade meningioma (World Health Organization [WHO] grade II and III) was significantly higher than that in low-grade ones (WHO grade I) (p = 0.002) and significantly correlated with the MIB-1 labelling (cell proliferation marker) index (r = 0.338, p = 0.009) and mitotic count of the tumour (r = 0.284, p = 0.03). The ROC analysis revealed that the TGR of 1.0 was the best cutoff value for detecting high-grade meningioma with 43% sensitivity, 95% specificity, and 81%accuracy.
Dual-phase metabolic 18F-FDG PET approach with subsequent quantitative analysis was undertaken by Ghany et al. [28] in order to discriminate grading of gliomas. They found good correlation between the dual-phase PET grading and the histopathological grading of gliomas. When a 23% increase was used as the cutoff for analysis of the difference in SUVmax of the lesion versus normal grey matter over time, sensitivity was 88.9%, specificity 85.7%, and accuracy 89.4% (p = 0.003; AUC = 0.94). Nakamura et al. [29] investigated the connection between quantitative features of 18F-FDG uptake by endometrial carcinomas and International Federation of Gynaecology and Obstetrics (FIGO) grade: they found significant correlations between the SUVmax of the primary tumour and the FIGO grade, maximum tumour size, and glucose transporter-1 expression. Furthermore, multivariate analysis showed that the FIGO grade of endometrial cancer was most significantly identified as a relation factor of SUVmax (≥ 17.6). Rakheja and Probst [30] studied 18F-FDG uptake parameters for grading sarcomas and concluded that while 18F-FDG PET/CT cannot replace histopathology in the diagnosis of sarcoma, it is certainly of use in guiding surgeons and pathologists to biopsy the most aggressive regions of the tumour.
Similar to the above-mentioned studies, current one investigates connection of various metabolic PET parameters to tumour grade, but the most promising results are achieved when multiparametric modelling, utilising both semiquantitative and textural parameters, is applied.
Several limitations of the current study should be acknowledged. First of all, it was based on a single institution dataset, collected from one PET/CT scanner. Secondly, distribution of different tumour grades in our dataset is rather unequal, with grade 1 tumours comprising roughly only 10% of the sample. Consequently, these results should be further validated in a multi-institutional or multi-scanner scenario, in order to test whether the suggested model will withstand the difference in values of quantitative parameters extracted from images generated with different reconstruction algorithms, and on a larger patient population.
In conclusion, multiparametric modelling with GMDH utilising both semiquantitative and texture quantitative metabolic PET parameters seems to be an interesting tool for noninvasive malignant epithelial tumours grade differentiation. Results achieved in our dataset allow for hypothesise that, despite difference in absolute values generated by segmentation techniques, relative differences in combination of multiple parameters inside the subsets from different segmentation techniques allow to correlate with different tumour grades. Among extracted features, conventional semiquantitative and volumetric parameters demonstrated significant dependence from segmentation technique, while three quantitative texture indices remained stable. This allows to speculate that metabolic image features responsible for reflecting difference in tumour biology and grade are possibly more likely to be represented by heterogeneity of tracer uptake, rather than its intensity. Further investigation is required with larger patient population in order to validate the potential value of this approach.